Research map / Image generation / Text-to-image diffusion
Correlation-Weighted Multi-Reward Optimization for Compositional Generation
Jungmyung Wi, Hyunsoo Kim, Donghyun Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add correlation-based reweighting (CR) (ablation) | CMO SD3.5-M · ConceptMix avg full mark | avg full mark: 0.3531 -> 0.3913 | Improved |
| add correlation-based reweighting (CR) (ablation) | CMO SD3.5-M · ConceptMix avg concept fraction | avg concept fraction: 0.7993 -> 0.8126 | Improved |
| add single-attribute training prompts (ablation) | CMO SD3.5-M · ConceptMix avg full mark | avg full mark: 0.2713 -> 0.3305 | Improved |
| add multi-concept training prompts (ablation) | CMO SD3.5-M · ConceptMix avg full mark | avg full mark: 0.3305 -> 0.3531 | Improved |
| apply DRO (MixGRPO + GDPO) alone on standard data vs base (ablation) | SD3.5-M · ConceptMix avg full mark | avg full mark: 0.2667 -> 0.2713 | Mixed |
| increase group size from 8 to 16 (ablation) | CMO SD3.5-M · ConceptMix avg full mark | avg full mark: 0.2934 -> 0.3919 | Improved |
| group size 4 instead of 16 (ablation) | CMO SD3.5-M · ConceptMix avg full mark | avg full mark: 0.3919 -> 0.0000 | Hurt |
| train FLUX.1-dev with CMO | FLUX.1-dev · ConceptMix full mark (k=7) | full mark score: 0.0372 -> 0.1883 | Improved |
| CMO vs Qwen-Image at k=7 | Qwen-Image · ConceptMix full mark (k=7) | full mark score: 0.1747 -> 0.1883 | Improved |
| CMO vs Qwen-Image on concept fraction at k=7 | Qwen-Image · ConceptMix concept fraction (k=7) | concept fraction score: 0.7942 -> 0.7884 | Hurt |
| train SD3.5 with CMO | SD3.5 · T2I-CompBench | Avg score: 0.5239 -> 0.6141 | Improved |
| train FLUX.1-dev with CMO | FLUX.1-dev · T2I-CompBench | Avg score: 0.4866 -> 0.5961 | Improved |
Closest papers by meaning
- ConceptMix: A Compositional Image Generation Benchmark with Controllable Difficulty (NeurIPS 2024 · similarity 0.88)
- HiCoGen: Hierarchical Compositional Text-to-Image Generation in Diffusion Models via Reinforcement Learning (CVPR 2026 · similarity 0.88)
- ReNO: Enhancing One-step Text-to-Image Models through Reward-based Noise Optimization (NeurIPS 2024 · similarity 0.87)
- Multimodal Prompt Optimization: Why Not Leverage Multiple Modalities for MLLMs (ICLR 2026 · similarity 0.86)
- IterComp: Iterative Composition-Aware Feedback Learning from Model Gallery for Text-to-Image Generation (ICLR 2025 · similarity 0.86)
- Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
